What the 2027 enterprise AI stack looks like

We build AI products for a living, which means we place bets on architecture every month and find out quickly when we are wrong. This piece is not a forecast of which model wins. It is a field report on the shape the enterprise stack is settling into, because that shape is already visible in what survives contact with production.

The layer that commoditises

Frontier models leapfrog each other every few months, and the striking thing is how little that now disrupts a well-built product. Capability differences that were chasms in 2023 are gaps in 2026, and for most enterprise workloads several providers now clear the bar. Prices per token keep falling. Open-weight models keep closing in from below.

When several suppliers can meet the spec and switching is cheap, you are describing a commodity. The consequence for architecture is blunt: any system that hard-wires itself to one model is storing up a migration. Every serious platform we build now treats the model as a swappable component behind an internal interface, chosen per task: a frontier model for judgement-heavy work, a fast cheap model for classification and routing, sometimes an open-weight model where data must not leave.

The layers that appreciate

If the intelligence is rented, what is owned? Four layers, and they are where the 2027 stack differentiates.

Orchestration. The machinery that decomposes an objective into steps, routes each step to the right tool or model, handles failure and retries, and knows when to stop and ask a human. This is where agentic products actually live. It encodes judgement about the work itself, which no model vendor can ship.

Memory and context. The organisation's documents, records, rules and history, structured so systems can use them safely. Context is the moat models cannot cross on their own: two companies calling the identical model get entirely different value depending on what it can see. The winners treat this as an asset with an owner, not exhaust.

Audit and control. Execution logs, approval gates, permission boundaries. In 2024 these were an enterprise checkbox. In 2027 they are the product surface that decides what an organisation will allow AI to touch, and therefore how much value it can capture. We now design the audit trail before the feature.

Evaluation. The quiet layer nobody demos: the harness that proves, continuously, that the system still does what it should as models underneath it change. Teams with real evaluation ship model upgrades in days. Teams without them freeze on old models out of fear, and fall behind at commodity speed.

What this means if you are buying

Three questions cut through most vendor theatre:

1. "Which models do you run on, and how do I switch?" A confident answer signals sound architecture. A single-model answer signals their migration risk is priced into your contract.
2. "Show me the log of what your system did yesterday." Products built for production show you immediately. Products built for demos change the subject.
3. "How do you evaluate a model upgrade before my users see it?" Silence here is the loudest signal of all.

What this means if you are building

Spend your scarce engineering on the layers that appreciate. Wrap the model, never marry it. Put orchestration, memory, audit and evaluation on your balance sheet as the assets they are. And keep the interface to intelligence thin, because everything below it will be different in eighteen months, and if you have built well, that will be good news rather than a rewrite.

The companies that internalise this will experience model progress as a tailwind. Everyone else will experience it as churn.